Paradox in AI – AI 2.0: the way to machine
consciousness
Peter Palensky1 , Dietmar Bruckner2 , Anna Tmej2 , and Tobias Deutsch2
1
University of Pretoria, South Africa,
[email protected]
2
Vienna University of Technology, Austria
{bruckner,tmej,deutsch}@ict.tuwien.ac.at
Abstract. Artificial Intelligence, the big promise of the last millennium,
has apparently made its way into our daily lives. Cell phones with speech
control, evolutionary computing in data mining or power grids, optimized
via neural network, show its applicability in industrial environments. The
original expectation of true intelligence and thinking machines lies still
ahead of us. Researchers are, however, optimistic as never before. This
paper tries to compare the views, challenges and approaches of several
disciplines: engineering, psychology, neuroscience, philosophy. It gives a
short introduction to Psychoanalysis, discusses the term consciousness,
social implications of intelligent machines, related theories, and expectations and shall serve as a starting point for first attempts of combining
these diverse thoughts.
Key words: Machine consciousness, artificial intelligence, psychoanalysis
1 Introduction
Embedded computer systems have seen their computing power increase dramatically, while at the same time miniaturization and wireless networking allow
embedded systems to be installed and powered with a minimum of support infrastructure. The result has been a vision of “ubiquitous computing”, where
computing capabilities are always available, extremely flexible and support people in their daily lives.
The common view in related communities is that computers will not only
become cheaper, smaller and more powerful, but they will disappear and hide
or become integrated in normal and everyday objects [21], [22]. Technology will
become invisible and embedded into our surroundings. Smart objects communicate, cooperate and virtually amalgamate without explicit user interaction or
commands; they form consortia for offering or even fulfilling tasks for a user.
They are capable of not only sensing values, but deriving context information
about the reasons, intentions, desires and beliefs of the user. This information
may be shared over networks – one of which is the world-wide available Internet
– and used to compare and classify activities, find connections to other people
2
Peter Palensky et al.
and/or devices, look up semantic databases and much more. The uninterrupted
information flow makes the world a global village and allows the user to access
his explicitly or implicitly posed queries anywhere anytime.
The vision of machine consciousness – which includes the availability of
enough computation resources and devices for fulfilling the tasks – poses many
requirements across the whole field of information and communication technology and allied fields [23]. For example, the development in ambient intelligence
with respect to conscious environments requires development in the areas of sensors, actuators, power supplies, communications technology, data encryption and
protection, privacy protection, data mining, artificial intelligence, probabilistic
pattern recognition, chip design and many others not stated here. Each research
group and even researcher has its own view on what machine consciousness
will be or will approximate. The same is already true for the many necessary
technologies required for machine consciousness. The research on ambient intelligence for instance can be divided up [24] into three basic research areas – or
existing research projects can be grouped into projects investigating three basic methods – which reflect fundamentally different approaches in establishing
ubiquitous computing environments: Augmented Reality (see for example [25]),
Intelligent Environments ( e.g. [26], [27], [28]), and Distributed Mobile Systems
(see [29], [30]).
The border between smart devices and devices with machine consciousness
may be something very controversial and from the philosophic point of view of
massive importance and implications. This will be true for ambient intelligence
applications as well as any other, e.g. industrial, conscious application. However,
for the user of such a device this distinction is meaningless: A user expects the
fulfillment of a particular task, the device just needs to be intelligent enough to
to so. An argument for calming down the apprehension of machines that will
control us is presented below with the model of nested feedback loops, each loop
representing a bit more consciousness than the lower ones. In this context the
designer of a machine can control the level of consciousness he wants from his
device.
These considerations together give a rough definition of machine consciousness as we see it: Machines equipped with decision units that allow them to
think in a way humans do. The key issue is the “thinking in a way humans do”
in opposition to “acting looking like humans” of many well-know projects, which
is nothing but mimicry. There are, e.g., robots capable of performing facial expression like smiling. However, this does not imply that the robot “is amused”
as a human would be who smiles, or that the robot has any other human-like
intentions to smile. But this would be the necessary requirement for attributing
this robot consciousness.
AI 2.0
3
2 What consciousness?
Consciousness1 helps us humans to put ourselves into the set of parameters and
actors when decisions are made. A concept of self, its desires and plans is one
attribute that makes someone appear intelligent.
Most of all, consciousness is a subjective quality. Person A can feel its own
consciousness, be aware of it, test it and have a concept of itself. Person A can
tell Person B about this extraordinary experience but Person B can never be
sure. Only from “inside”, consciousness can be experienced and verified. The
link of this qualia to the physical process, the physical machine, is unfortunately
not clear [35].
The outer shell of a conscious being might exhibit very distinctive behavioral patterns. These patterns might in turn be checked against a “turing test on
consciousness”. There are, however, numerous arguments about machines, potentially passing these tests but still being “zombies”, i.e. lacking consciousness
[36].
Another distinction is where we assume consciousness. Generally accepted as
a human phenomenon (maybe also in “higher” mammals), its projection to or
implementation in machines [37] opens up two problems:
1.) We could be fooled by anthropomorphic aesthetics. Humans actively seek
emotions in faces, even if they are inanimate rubber masks and we should be
aware of this pitfall.
2.) If we created this machine, we have total power over its hardware (and
software, if we use contemporary terms) and can copy it, switch it off and on
without damage, modify it, monitor it without interference, etc.
It is the second point that would make these beings massively different to
humans. We have no power over our own hardware. We are mortal and do not
fully understand how our body works. Out of this situation we can derive two
alternative outcomes:
a.) A machine that can be switched off and on cannot host consciousness
because it would be far too primitive. Once we have machine consciousness we
will realize that the hardware has been gradually taken out of our hands (e.g.
manufactured by nano-bots and evolutionarily modified beyond our knowledge)
and therefore as intangible as our own brains.
or
b.) It works and constitutes one half of what many people dream of: a potential “storage” or platform for our own consciousness, ultimately leading to
immortality (the other half, however, is still missing: understanding of our own
hardware a.k.a. “wetware”, to download its content to the new platform). Unfortunately, a subjective engineering estimate would be that the chances that
such a machine – if it develops something like consciousness – is compatible to
our “software” are virtually zero.
1
The interested reader might be referred to David Chalmers’ and David Bourget’s web
repository http://consc.net/online, which lists more than five thousand papers on
consciousness.
4
Peter Palensky et al.
So we might reach some point where we have remarkable and potentially
conscious artifacts, but their comparison to our own body and mind as a unity
is very questionable, although most technical applications would be happy with
the mind of a sophisticated “zombie” [42].
3 Consciousness support
To provide a machine with machine consciousness one must, in order to stay scientifically consistent, review already existing research on human consciousness.
In this chapter we will give an overview about how various sciences approach
consciousness. For technical purposes a holistic, functional model without contradictions is desired. We will choose an existing method as a template and go
into more depth into that theory.
3.1 Philosophy
The questions of consciousness that philosophy traditionally dealt with can be
gathered in three crude rubics as the What, How and Why questions: What
is consciousness, its principal features? How does consciousness come to exist,
and finally, Why does consciousness exist? David Chalmers [7] summarizes more
modern philosophical approaches when he differentiates between the so-called
“hard problem” and the “easy problem” of consciousness. The latter concerns
objective mechanisms of the cognitive systems: the discrimination of and reaction to sensory stimuli, the integration of information from different sources and
the use of this information to control behavior and the verbalization of internal
states, while it is the “hard problem” that actually deals with the “mystery”
of consciousness [7], p. 62: the question of how physical processes in the brain
give rise to subjective experience. This involves the inner aspect of thought and
perception: the way things feel for the subject. This part of consciousness is
also called “phenomenal consciousness” and “qualia” [8], while Chalmers’ “easy
problem” is also called “access-consciousness”. Daniel Dennett [9], on the other
hand, denies that there is a “hard problem”, asserting that the totality of consciousness can be understood in terms of impact through behavior. He coined
the term “heterophenomenology” to describe an explicitly third-person scientific
approach to human consciousness.
Dennett’s new term is closely linked to his cognitive model of consciousness,
the Multiple Drafts Model (MDM) [9]. According to this model, there are a variety of sensory inputs from a given event and also a variety of interpretations
of these inputs. The sensory inputs arrive in the brain and are interpreted at
different times, so a given event can give rise to a sequence of discrimination,
constituting the equivalent of multiple drafts of a story. As soon as each discrimination is accomplished, it becomes available for eliciting a behavior. Like
a number of other theories, the Multiple Drafts Model understands conscious
experience as taking time to occur. The distinction is that Dennett’s theory
AI 2.0
5
denies any clear and unambiguous boundary separating conscious experiences
from all other processing. According to Dennett, consciousness is to be found
in the actions and flows of information from place to place, rather than some
singular view containing our experience. The conscious self is taken to exist as an
abstraction visible at the level of the intentional stance, akin to a body of mass
having a center of gravity. Similarly, Dennett refers to the self as the center of
narrative gravity, a story we tell ourselves about our experiences. Consciousness
exists, but not independently of behavior and behavioral disposition, which can
be studied through heterophenomenology.
3.2 Psychology
Before the advent of cognitive psychology, psychology failed to satisfactorily
study consciousness. Introspectionism, first used by Wilhelm Wundt as a way
to dissect the mind into its basic elements, dealt with consciousness by selfobservation of conscious inner thoughts, desires and sensations. The relation of
consciousness to the brain remained very much a mystery. This experimental
method was criticized by upcoming behaviorists as being unreliable; scientific
psychology should only deal with operationalizable, objectifiable and measurable contents. Behaviorism thus studied the mind from a black-box-point-ofview, stating that the mind could only be fully understood once the inputs and
the outputs were well defined, without even hoping to fully understand the underlying structure, mechanisms, and dynamics of the mind.
Since the 1960s, cognitive psychology has begun to examine the relationship
between consciousness and the brain or nervous system. The question cognitive
psychology examines is the mutual interaction between consciousness and brain
states or neural processes. However, despite the renewed emphasis on explaining
cognitive capacities such as memory, perception and language comprehension
with an emphasis on information processing and the modeling of internal mental
processes, consciousness remained a largely neglected topic until the 1980s and
90s.
A major example of a modern cognitive approach to consciousness research
is the global workspace theory of Bernard Baars [1]. It offers a largely functional
model of consciousness which deals most directly with the access notion of consciousness and has much in common with the multiple drafts model mentioned
above. The main idea of global workspace theories is that consciousness is a
limited resource capacity or module that enables information to be “broadcast”
widely throughout the system and allows for more flexible sophisticated processing. It is thus closely allied with many models in cognitive psychology concerned
with attention and working memory.
Just as the philosophical approaches mentioned above, however, psychological attempts at describing and investigating consciousness lack a thorough and
exact analysis of the structure, mechanisms and dynamics of consciousness and
mental processes.
6
Peter Palensky et al.
3.3 Evolutionary Biology
Evolutionary Biology has mainly investigated the question of the causes for
consciousness. From this point of view, consciousness is viewed as an adaption
because it is a trait that increases fitness.
3.4 Physics
Modern physical theories of consciousness can be divided into three types: theories to explain behavior and access consciousness, theories to explain phenomenal
consciousness and theories to explain the quantum mechanical (QM) Quantum
mind [43], [44]. These latter theories are based on the premise that quantum mechanics is necessary to fully understand the mind and brain to explain consciousness. The quantum mind hypothesis proposes that classical mechanics cannot
fully explain consciousness and suggests that quantum mechanical phenomena
such as quantum entanglement and superposition may play an important part in
the brain’s function and could form the basis of an explanation of consciousness.
3.5 Cognitive Neuroscience
This scientific branch is the most modern approach to consciousness research,
primarily concerned with the scientific study of biological substrates underlying cognition, with a specific focus on the neural substrates of mental processes
and their behavioral manifestations. Amongst other things, it investigates the
question of how mental processes uniquely associated with consciousness can
be identified. It is based on psychological statistical studies and case studies of
consciousness states and the deficits caused by lesions, stroke, injury, or surgery
that disrupt the normal functioning of human senses and cognition. Cognitive
neuroscience is a branch of both psychology and neuroscience, unifying and overlapping with several sub-disciplines such as cognitive psychology, psychobiology
and neurobiology. One major question that cognitive neuroscience deals with is
the so-called “mind-body-problem” [17], the question how brain and mind or
brain and consciousness relate to each other. Many cognitive scientists today
hold the view that the mind is an emergent property of the brain: mind and
brain equally exist, however they exist at different levels of complexity.
Antonio Damasio [2] differentiates between “core consciousness” and “extended consciousness”. While core consciousness describes a hypothesized level
of awareness facilitated by neural structures of most animals, which allows them
to be aware of and react to their environment, extended consciousness is a much
more complex form of consciousness, allowing for a sense of identity and personality and meta-consciousness, and linking past, present and future. Higher forms of
extended consciousness only exist with humans and depend on (working) memory, thinking and language. Just as all other scientific approaches mentioned
above, cognitive neuroscience, too, does not provide us with a detailed and concrete model of the mental processes involved in the structure and mechanisms
of consciousness.
AI 2.0
7
3.6 Psychoanalysis
Psychoanalysis opened up the research field of consciousness to the unconscious
properties of the mind. As described in more detail below, Freud developed his
metapsychological ideas from a theory of three mental processes – conscious,
preconscious and unconscious – to a theory of three agencies – the Ego, the Id
and the Super-Ego – functioning on the basis of these processes. Although Freud
developed his theory as a contrast to the psychology of consciousness exclusively
current at his time [10], naturally, the mind that Freud set out to analyze and
describe with great precision includes and determines consciousness: although
many thoughts or other psychic/mental contents may never reach consciousness,
they will always exert influence on consciousness and behavior. According to
Freud, it is therefore both legitimate and necessary to include those properties
that lie behind consciousness within our conception of the mind [17].
Freud and other psychoanalysts after him saw and see consciousness largely
as a means to perceive outer and especially also inner events and inner-psychic
qualities [31], as a property of the mind as opposed to the mind itself [17]. The
mind, the mental apparatus, including conscious and unconscious properties was
described by Freud in all its functions and dynamics. In the end, psychoanalysis
emerges as the only science dealing with human consciousness on a detailed
enough level to be used for an implementation of machine consciousness.
4 Psychoanalysis, the template
For the scope of this work psychoanalysis was chosen as the template theory of
human consciousness. The main reasons therefore are the functional approach of
modelling human thinking, which very well fits a computer engineering approach,
and the somewhat technical or natural scientific approach of Freud to the topic
which resulted in texts and arguments that can be followed and agreed on by
natural sceintists. This chapter gives an introduction to people not from the
field.
4.1 Basics
Psychoanalysis was founded by Sigmund Freud (1856-1939), originally a neurologist and neuroanatomist [6], who in his study of hysteria developed his first
ideas about unconscious affects, psychic energy and the cathartic effect of verbal
expression. In the course of his life, Freud developed these and other concepts
further, abolished some, modified others, while inspiring many other scientists
to join him in his quest. Some of these came to disagree with Freud in time and
went on to pursue their own strands of theory.
Freud himself [12] described psychoanalysis as follows: “Psycho-analysis is
the name (1) of a procedure for the investigation of mental processes which
are almost inaccessible in any other way, (2) of a method (based upon that
investigation) for the treatment of neurotic disorders and (3) of a collection of
8
Peter Palensky et al.
psychological information obtained along those lines, which is gradually being
accumulated into a new scientific discipline” (p. 235). He goes on to differentiate:
“The assumption that there are unconscious mental processes, the recognition
of the theory of resistance and repression, the appreciation of the importance
of sexuality and the Oedipus complex – these constitute the principal subjectmatter of psycho-analysis and the foundations of its theory” (p. 250).
After Freud died in 1939, psychoanalytic theory and practice continued to
be further developed by scientists and practitioners as Heinz Hartmann, Anna
Freud (Sigmund Freud’s youngest daughter), and Melanie Klein, to name just
very few. However, already before Freud’s death, there had been disagreements
about the meaning of certain concepts or ways of treatment. Since then, controversy has continued to be characteristic of psychoanalytic theory; distinct schools
of thought focusing on different topics, e.g. Ego-psychology or object relations
have emerged. Even today, more than 100 years after its foundation, psychoanalysis is a living science in the sense that the different concepts of the different
schools are still being further developed and discussed. While psychoanalysis
is for a great part concerned with psychotherapeutic methods, psychopathologies and individual developments leading thereto, one great part also consists of
metapsychology, or conceptualizing the psychic apparatus and the way it (mal)functions. As this part of the psychoanalytic theory shall be our main focus,
this introduction will simply exclude the other, albeit certainly very important
and also characteristic aspects of psychoanalysis.
Additionally, although, as described above, there are many different strands
of theory in psychoanalysis, each with different main focuses and different understandings of certain concepts, in this introduction, we will still concentrate
on Freud’s original conception of psychoanalytic theory. This is due to the very
fundamental and basal nature of Freud’s original ideas which to this day have
not been abolished, and continue to be valid also in modern psychoanalytic theory, although different schools may focus on different concepts from different
periods in Freud’s life and development of the theory. Details about the technical conceptualization and partly implementation thereof can be found in [40].
The authors there in a first attempt concentrated on employing the original
psychoanalytic theory, before venturing further to include other, more modern
psychoanalytic concepts, be it the case that these prove as usable and expedient
as Freud’s conception of the psychic apparatus.
4.2 The mental apparatus (1): the topographical model
Freud conceptualized the mind as being divided up into different parts, each the
home of specific psychological functions. His first model, the topographical model
[14], divides the mind into the unconscious, the preconscious and the conscious
system. Characteristic of these different parts are two different principles of
mental functioning [11] – the so-called primary and secondary processes. While
secondary process thinking – typical of conscious processes – is rational and
follows the ordinary laws of logic, time and space, primary process thinking
– typically unconscious – is characteristic of dreaming, fantasy, and infantile
AI 2.0
9
life in which the laws of time and space and the distinction between opposites
do not apply. Some psychological processes are however only unconscious in
the descriptive sense, meaning that the individual is not aware of them, but
they are easily brought to mind (preconscious). With dynamically unconscious
processes, on the other hand, it is not by simple means of effort or change of
attention that they can be rendered conscious. These to the conscious system
unacceptable psychic contents are subject to repression and operate under the
sway of primary processes.
4.3 Drive theory
Freud saw the internal world as dominated by man’s struggle with his instincts
or drives. In his initial formulation of instincts, Freud [15] distinguished between
self-preservative (e.g. hunger) and sexual drives (libido). Later, he stressed the
difference between sexual drives and aggressive or destructive drives (Eros vs.
Thanatos) [16]. Classically, instinctual wishes have a source, an aim, and an object. Usually, the source of the drive is infantile and lies in the body, possibly
in an erogenous zone. Over time, after several similar (real or imagined) satisfactions of instinctual wishes, source, aim and object begin to mesh together
into a complex interactional fantasy, part of which is represented in the system
unconscious.
4.4 The mental apparatus (2): the structural model
In the structural model, Freud [12] proposed three parts or structural components of the human mind: Id, Ego and Super-Ego.
Id
The Id is the first psychic structure of the mental apparatus, out of which, in
the course of infantile and childhood development, the Ego and the Super-Ego
evolve. It contains the basic inborn drives and sexual and aggressive impulses, or
their representatives. As such, it is an inexhaustible source of psychic energy for
the psychic apparatus: The Id’s wishes strive for immediate satisfaction (pleasure
principle) and therefore drive the functions of the Ego to act. Its contents are
unconscious and function following the primary process.
Ego
In the course of infantile development, the perceptive and executive parts of
the Id, responsible for drive satisfaction by perceiving the baby’s environment
and ways to gain satisfaction, start to form a new part of the mental apparatus:
the Ego. The mature Ego’s tasks, however, exceed perception and execution: the
Ego has to control the primitive impulses of the Id and to adapt these to outer
reality (reality principle) as well as to mollify the requirements of the Super-Ego.
For these purposes, the Ego makes use of so-called defense mechanisms such
as repression to keep unacceptable impulses within the Id and therefore evade
conflict with either outer reality or Super-Ego-requirements. The contents of the
Ego are partly conscious, partly unconscious. A however certainly not complete
10
Peter Palensky et al.
list of Ego functions should not omit the following [5]: consciousness; sensory
perception; perception and expression of psychic agitation; thinking; controlling
motor functions; memory; speech; defense mechanisms and defense in general;
fighting, controlling and binding drive energy; integrating and harmonizing and
reality check.
Super-Ego
The Super-Ego comprises the conscience and ideals, thus allocating (moral)
rules and prohibitions which are derived through internalization of parental or
other authority figures, and cultural influences from childhood onwards. The
Super-Ego resembles the Ego in that some of its elements are easily accessible
for consciousness, while others are not. Super-Ego ideation also comprises rational and mature presentations up to very primitive and infantile ones. The task
of the Super-Ego is to impact on the actions of the Ego, especially to support
it in its defensive actions against the drives with its own moral rules. However,
the relationship between the Ego and the Super-Ego will not always be this harmonic: in other cases, e.g. if the difference between instinctual or other repressed
wishes from the Id and moral rules from the Super-Ego becomes to great, the
Super-Ego produces feelings of guilt or a need for punishment inside the Ego.
4.5 Psychoanalysis and Neuroscience
In recent years, a new scientific research strand has developed aiming at supporting and reassigning psychoanalytic concepts by neuroscientific findings: neuropsychoanalysis. Solms [18] provides a summary of the neuroscientific support
for some basic psychoanalytic concepts, one of which is the notion that most
mental processes occur unconsciously. Different memory systems have been identified, some of which are unconscious, which mediate emotional learning. The
hippocampuses, which lay down memories that are consciously accessible, are
not involved in such processes. Therefore, no conscious memories are available
and current events can “only” trigger remembrances of emotionally important
memories. This causes conscious feelings, while the memory of the past event
remains unconscious. In other words, one consciously experiences feelings but
without conscious access to the event that triggers these feelings.
The major brain structures for forming conscious memories do not function
within the first two years of life. Developmental neurobiologists largely agree
that early experiences, especially between infant and mother, fundamentally
shape our future personality and mental health. Yet none of these experiences
can be consciously remembered. It becomes increasingly clear, that a great deal
of our mental activity is unconsciously motivated.
Freud’s ideas regarding dreams – being instigated by the drives and expressing unconscious wishes – were at first discredited when rapid-eye-movement
sleep and its strong correlations with dreaming were discovered. REM sleep
occurred automatically and was driven by acetylcholine produced in a “mindless” part of the brain stem, which has nothing to do with emotion or motivation. Dreams now were regarded as meaningless, simple stories concocted by
the brain under the influence of random activity caused by the brainstem. Yet
AI 2.0
11
recent work has revealed that dreaming and REM sleep are dissociable while
dreaming seems to be generated by a network of structures centered in the forebrains instinctual-motivational circuits. These more recent views are strongly
reminiscent of Freud’s dream theory.
These developments and the advent of methods and technologies (e.g.: neuroimaging) unimaginable 100 years ago make it possible today to correlate the
psychoanalytic concepts, derived from observation and interpretation of subjective experiences, with the observations and interpretations of objective aspects
of the mind studied by the classical neurosciences, thus rendering psychoanalysis
and psychoanalytic theories objectifiable.
5 Road-map
The capabilities of the desired machine consciousness introduced above represent
a development stage in the future, which requires several disruptive innovations.
We cannot say today how long it will take to reach them. There are many other
authors predicting human-like intelligence within some decades, or only in 100
years, or even never. The authors here believe we are much closer! We see many
promising approaches to problems of AI, their only drawback lies in their position
within the framework of the application – which, in our opinion, shall be given
by psychoanalysis [40]. However, psychoanalysis is not concerned with all the
necessary functions, it is just and primarily concerned with psychic functioning
itself.
Psychoanalysis requires so-called memory traces. A memory trace is the mental representation of something, be it a person, an object, event or whatever. Psychoanalysis is not concerned with perception (which generates memory traces),
or with direct motor control. For technical systems these are however key requirements, needless to say these are the fields where robotics, AI (e.g. Software
Agents) and others put most of their efforts in, but we need to search for other
templates than psychoanalysis. Candidates here are developmental psychology
and neuroscience, to mention just two. It becomes more and more clear that
a system showing something like machine consciousness needs to be designed
following a model, of which the mentioned disciplines have different viewpoints.
Therefore, it needs to combine them in a holistic manner.
Psychoanalysis together with the other mentioned fields span a huge search
space for engineers. Therefore it is not possible (and useful) to start trying to
implement all of it. It is necessary to define a road map with milestones and
priorities. A good hint is given by Rodney Brooks in his recent Spectrum article
[34], where he introduced four capabilities of children, which are vital behavioristic observable capabilities of human-like intelligent machines. Unfortunately,
these have nothing to do with psychic functioning; they are just observations of
behavior. There may however be one connection in the last point – the “theory of
mind” which in psychoanalysis is called “mentalization” or “reflective functioning” [39], which is definitely a crucial capacity within AI (being able to guess
12
Peter Palensky et al.
what another person is thinking, planning; understanding why a person does
whatever it is they are doing etc.).
5.1 Mental functions
Before mentioning and commenting Brooks’ list, it is necessary to explain some
of the very basic concepts of psychoanalysis that need to be addressed by the
first machines with comparable behavior. These are the primary and secondary
process, terms that refer to the quality of thoughts – or mental content in terms
of uncontrolled, impulsive actions or considered ones. The functions necessary
for this according to [19] are:
perception: The mind knows two kinds of perception, internal and external.
Internal perception results from an observational perspective on the mind.
Both together allow for mental content of the kind: “I am experiencing this”.
memory: The mind recognizes previous mental experiences of the mentioned
kind. It is also able to derive cause-and-effect sequences thereof. These two
capabilities together form the immature ego, which only depends on drives
and environment.
emotions: The very basic function of emotion is to rate something in terms of
good or bad in a biological sense. More biologically successful actions are
felt satisfying. Emotions are used to rate the above mentioned perceptions
and memory content. In this way, quantitative events acquire quality.
feelings: Feelings are the base for consciousness. Their function is to evaluate
the other mechanisms. The kind of mental content that results is “I feel
like this about that”. Feelings are also stored together with their generating
mental experiences. They can be used to create motivation by trying to
repeat previous experiences of satisfaction. Motivation searches through past
experiences and matches them with the present to come to decisions about
what is to be done. The decisions lead to actions (motor output). These all
together form the primary process.
inhibition: Experience about actions show that some of them are only satisfying in the short-term, but not in the long-term. Therefore it is necessary to
be able to tolerate temporary unpleasure via inhibiting the immediate action plan. This capacity permits thinking. Thinking is inhibited (imaginary)
action, which permits evaluation of potential (imaginary) output, notwithstanding current (actual) feelings. This function is called secondary process.
It replaces immediate actions with considered ones.
The ability of the secondary process develops very early in a child at the age
of around two. However, the finesse and long-term anticipation capabilities in
imaginary actions develops during the whole life.
The above description lets us conclude that the mind is organized as a system
of nested feedback loops. If we leave memory aside, the first level consists of
perception and evaluation. Let’s call the pure perception mental content level
AI 2.0
13
12 . The evaluation together with the perception, towards it is targeted, are again
stored. This mental content is another piece of information; let’s call it mental
content level 2. Content of level 2 is rated with feelings. These together form
mental content level 3. The concept of evaluating mental content with feelings
from level 3 upwards is the same for the all upper layers. It is not clear at this
point if there are intermediate levels, but one of the next levels is related to
language. With this, some mental content reaches a level of attention which is
important enough to give it a name or a phrase – and therefore being able to
share it with others. Again, in the upper levels, reasoning and feeling about
mental content related with language – dialogs, names, etc. – form the next
levels.
It is important that in this concept the lower levels of some mental content
need to be “created” before upper levels can even be thought of. This may be
one reason why it is so hard to bring findings or knowledge of different context
together.
5.2 Example
As an example let us think about the following situation: Somebody (being it a
child, an ancient hunter, whoever) is walking in nature and sees a steep hillside,
an edge. Mental content level 1 would be “I see an edge”. For some reason (maybe
searching for food, playing, etc.) this edge generates the interest of the person,
letting it think (conscious or not) “I am interested in that edge”. Therefore, the
person moves itself to that place. This involves basic action planning. However,
when the person reaches the edge, the region behind gets visible. Let us further
assume the person sees something he has searched for. Then mental content of
the kind “I like this place – and also the landmark (the edge) which indicates the
place” could arise, because walking there results in reaching something desired.
If this desired thing is interesting or important enough, the person could think
“I would like to tell my spouse of this” and “I call this region (e.g.) the XYZ
place and the landmark the RST edge”.
In this way, perception is enhanced with evaluation level by level. The more
interest (positive or negative) it creates, the more chances for further processing
are implied. Things which cannot be perceived (because of a lacking template)
or which generate neutral evaluation are considered unimportant.
In the following the four capabilities mentioned by Brooks are described
(the first statements of the following sections marked with quotes). We try to
formulate necessary psychic functions related to the introduced levels of mental
content that enable the observation of the desired behavior. Additionally, other
necessary concepts for rudimentary machine consciousness beyond Brooks’ ones
are introduced.
2
Here we need to mention that “perception” itself requires lots of computation before
generating the above mentioned memory traces. So, in a holistic model, we would
start with level 1 on the sensor level.
14
Peter Palensky et al.
5.3 The object-recognition capabilities of a 2–year–old child
“A 2–year–old can observe a variety of objects of some type – different kinds of
shoes, say – andsuccessfully categorize them as shoes, even if he or she has never
seen soccer cleats or suede oxfords. Today’s best computer vision systems still
make mistakes – both false positives and false negatives – that no child makes.”
Children recognize shoes, not their size, weight, etc. They recognize shoes because of their roughly similar shape and their functionality. That means: children
know what shoes in general look like (foot-shaped plus a place where the foot
can be inserted) and what they are for (that is: for putting them on and walking
around). The rest is variable: exact shape, color, price; and does not interfere
with basic shoe-recognition. Therefore, also a giant, two-meter shoe placed on
the top of the roof of a giant shoe-shop will be recognized as a shoe, although
it technically is, of course, not a shoe (but: it is foot-shaped, even though giant,
and you could theoretically put your foot into it and therefore also walk around
in it); however the child is naive enough not to know about other things that
would hinder this shoe from really functioning in a shoe-like manner.
In visual object recognition requirements this would translate into many
problems, an algorithm would need to solve: Finding out the geometry of the
picture. How large is the area covered, how large are the objects. There needs
also to be a model of the object in 3D. The visual classificator has to be based not
only on features representing the spatial change in e.g. luminosity of pattern, but
needs to take contours into consideration. A promising approach in this respect
could be a combination of the boosted cascade of feature detection [20] relying
on a set of features like shape, color, etc., not only pattern, as described e.g. by
[41].
The very important aspect however is deriving the meaning of things for the
machine. What can the machine do with that object? Can it be evaluated good
or bad (or some more categories) in terms of the application of the machine?
5.4 The language capabilities of a 4–year–old child
“By age 4, children can engage in a dialog using complete clauses and can handle
irregularities, idiomatic expressions, a vast array of accents, noisy environments,
incomplete utterances, and interjections, and they can even correct nonnative
speakers, inferring what was really meant in an ungrammatical utterance and
reformatting it. Most of these capabilities are still hard or impossible for computers. ”
As elaborated above, the meaning is the essential thing. Language is a data
transmission channel to allow the creation of mental content in the mind of
the listener as desired by the speaker. It is clear that the rules of grammar and
language (word types, irregularities, categories, etc.) need to be learned (stored),
however, language represents another level of quality for mental experiences
in terms of perception, memory, emotions, and feelings. A lot of learning and
personally evaluating of lower level mental content is involved before.
AI 2.0
15
5.5 The manual dexterity of a 6–year–old child
“At 6 years old, children can grasp objects they have not seen before; manipulate
flexible objects in tasks like tying shoelaces; pick up flat, thin objects like playing
cards or pieces of paper from a tabletop; and manipulate unknown objects in
their pockets or in a bag into which they can’t see. Today’s robots can at most
do any one of these things for some very particular object. ”
The addressed capabilities are closely related to imagination. Imagination
is the reasoning about imaginary actions. It requires a large storage pool of
already perceived actions and thoughts together with their evaluation in various
respects. On the other hand, those capabilities require perfect motor control –
which is something where machines can finally (in the context of this paper)
compete with or even outperform humans.
5.6 The social understanding of an 8–year–old child
“By the age of 8, a child can understand the difference between what he or
she knows about a situation and what another person could have observed and
therefore could know. The child has what is called a “theory of the mind” of the
other person. For example, suppose a child sees her mother placing a chocolate
bar inside a drawer. The mother walks away, and the child’s brother comes and
takes the chocolate. The child knows that in her mother’s mind the chocolate
is still in the drawer. This ability requires a level of perception across many
domains that no AI system has at the moment.”
Actually, this is the only point really in connection with psychoanalysis – as
a theory about the adult mind. (see above), and being able to “mentalize” or
having a “theory of mind” really is a crucially important capacity for any social
interaction to succeed: we always consider the other’s point of view: what they
might be thinking or feeling at the moment (always being aware of the fact that
we will never be able to fully know what it is they are thinking or feeling), which
mental states make them behave in a certain way, or considering that they do
not know what we know. This capacity works in the background and is certainly
not consciously intended, but still it is there and absolutely necessary for our
understanding of ourselves, other persons and the interactions we come upon.
6 Necessary basic concepts
The above depicted road-map takes some concepts as granted. Among them are
embodiment, a complex environment and a social system. This section gives a
short overview on these topics.
Following the concept of “embodied cognitive science” as described in [32],
intelligence must have a body that again shapes the way an agent is able to think.
The body can be clearly distinguished between the surroundings, although the
body has to contain an interface that grants bidirectional communication to the
16
Peter Palensky et al.
surroundings. Therefore, the body contains different sensors to sense the surroundings of the agent. Compared to the human’s body, the sensors take over
the functionality of our five senses (taste, sound, tactile, sense of smell and vision) but have, due to the duty of the agent other functionality. To directly
interact with the environment and fulfill the requirement of proactiveness of an
autonomous agent, it also has to be equipped with actuators. The more complex
the actuators are, the higher is the degree of freedom, the agent can interact and
the more possibilities exist to reach desired goals. To lend importance to the
body, it is equipped with internal sensors, that are monitoring internal values
that are, together responsible for the homeostasis of the agent. This can be in
terms of robotic agents e.g. energy level, processor load or internal network load,
responsible for fast and slow internal message systems, comparable to the humans hormone system. Since, an agent shall be able to learn, also direct feedback
through the internal sensors is eligible as a result of taken actions. Therefore, the
environment has to contain dangerous components to gain deeper insight into
its own body and allows approximating limits that are given in the environment.
Each agent is placed into an ecological niche – defined by the environment and
the possibilities of the agent to interact with it. According to Brooks [33], intelligence is within the environment and not within a problem solver. Thereafter,
an agent is as intelligent as manifolded and/or complex its senses and actuators
are.
An environment in which agents operate has to contain enough distinguishable objects and areas to enable the agent to navigate within it. Like watchdogs
have good understanding where home and where outside is, software agents could
then develop similar abilities. Also emotional cathexis of objects leads in a complex world to better results. If only few different obstacles are available, almost
every place will have the same carthexis. The home will then be very positive,
areas close to enemies are emotionally difficult.
The social understanding of an 8 year old as depicted above is only a small
fraction of the complex social structures present in the human world. With
added complexity and intelligence of (multi-) agent systems dynamic societies
are needed. They are less complex as human ones, but nevertheless comparable
to them. A society needs social rules (predefined or emergent). Social rules can
be help others, don’t go there, etc. they glue the society together. They are also
some kind of reward system – he who owns great social reputation may receive
more help and has more possibilities to influence others. Different societies have
different characteristics: some value the experience of the elders, some the creative approaches, some strong leaders, etc. This is dependent on the needs of the
society. A further advantage of societies and their social rules and social system
is the possibility to specialize. One agent on its own has to do everything, in
combination with others he can specialize into one task. This could also be done
by design of a special task agent, but the specialization of an all purpose agent
into a few tasks has the advantage that if other agents fail to deliver the required
product/task (which has been agreed upon via the social system) the agent can
do it by himself – albeit in lower quality or slower.
AI 2.0
17
7 Implementation issues
Implementing traditional AI demands for mathematical skills, using optimized
algorithms and data structures. The above concepts, however, are very heterogeneous and have a rich structure. We have no reason to believe that an implicit
architecture might be trained to show the desired behavior. Therefore the possibility of an explicit implementation shall be assumed here. Explicit in this
context means that the various functional components and mechanisms, as described above, find their direct manifestation in a technical artifact – let’s assume
functional code.
Theoretically any technical function, described by a formal language, can be
implemented solely in hardware – or entirely in software. This assumption holds,
as long as we move withing the von-Neumann world of machines. Taken as a black
box, no-one cares about the engineering principles inside, as long as it behaves
on its interfaces as specified. Generating a specific mathematical function might
be done via a lookup-table or via a formula, while both can be implemented and
executed both in hardware or software (executed on an abstract, universal machine). The level of abstraction that the implementation shows towards the real
meaning is sometimes called the “semantic gap”, discussed in [4]. Its consequence
is decreased performance. The larger the gap, the worse the performance. This
derated performance can take extreme dimensions when the problem contains
chaotic aspects. The human mind in general show extremely complex, nonlinear
and chaotic behavior, stemming from incredibly large amounts of weighted memories that take influence on decisions. This exposed chaos gives an impression
on the complexity of the underlying information processing. An implementation
will face the same complexity and therefore strongly benefit from an as small as
possible semantic gap. The following “software-near” hardware are examples for
typical building blocks used in “small-gap machines”:
–
–
–
–
–
–
Associative memory with database-like capabilities in hardware
HW support for matrix operations
Dynamically adaptive microcode/hardware
Non-blocking multi-port memory
Merging of memory and operations, hardware-in-memory
Multidimensional, and not linear, connection of data and functions
All this is of course a violation of the pure von-Neumann idea of an abstract
machine that outsources all complexity and functionality into time and software
code. As current computers are (by implementing aspects of the above list)
already far away from this pure idea, this philosophical aspect should not bother
us.
In general, the above ideas of an intelligent machine demand a high level
of parallelism in software and subsequently in hardware, if we want operations
optimized. Parallelism, in turn, is a challenge to our technology, captivated in the
boundaries of space and time. 3D-silicon might be one step more but the level of
parallelism we need for instantaneously and successfully querying a tremendously
large memory reaches beyond that.
18
Peter Palensky et al.
Generally we can expect two sources of innovation that might bring us to the
stage to implement our abstract concepts
– Technological innovation
– Structural innovation
Even nowadays and throughout history these two phenomena have – typically
in alternating roles – brought technology forward. Just think of the development
of computers where technological progress like relais, tubes, TTL, CMOS etc.
were – usually when the current technology was on its limits – interlaced with
structural innovations like parallelism, superscalar architectures or multi-cores.
We can imagine massively parallel description languages that manifest in molecular nanotech and other currently unbelievable technologies to be the platform
of tomorrow. Structuring the functional blocks of our models, however, will still
be necessary. Taking X Billions of neurons does not create a human brain. It is
the structural order which makes it special.
One open question is the scalability of this concept. Implicit AI typically
has good scalability characteristics. A distributed genetic algorithm might for
instance be scaled via its population and genome size. Similar things can be
said of artificial neural networks, but as discussed in [3], neural networks do
unfortunately not surprise us with unexpected smartness if we allow for growing.
The capabilities of ANNs do not really aggregate. If “more” is wanted, the
ANN must show some macroscopical structure. The above concept’s structure
is, however, macroscopically not scaling. It does not necessarily get twice as
smart just if you introduce two semantic memories than just one. The size of the
individual components, especially filter networks or memories, surely can grow
if the given problem demands it.
Scalability plays a second role when thinking about really intelligent machines. “Singularity” evangelists (see the June 2008 issue of the IEEE Spectrum
for an excellent pro-con discussion on this topic) base their positive expectations on Moore’s Law and the daily experienced ever accelerating progress of
technology. Ironically, also critics use the economy of scale as argument. Nothing in nature (energy, matter, etc.) can grow forever, and exponential growth of
progress is – in their view – a subjective and potentially false perception. Even
if our “truly intelligent” machines might not be the paradise host for our consciousness, once it moves out of its current medium, they might be attributed
conscious attributes.
8 Social implications
The relation between truly intelligent machines and mankind is subject of numerous science fiction stories, the murderous and tragic ones dominating. Isaac
Asimov created a series of such stories, funny enough it is always man who
fails and causes trouble, and never machines who are described honest and innocent. We should not underestimate the mental abilities of man to cope with
AI 2.0
19
this challenge when complex machines are all around. We can expect a pretty
sober reaction by people that grow up with mass media and ubiquitous communication. These things will have their place in our world. More interesting is the
question what place we will have in their world?
Generally we should distinguish between three different encounters that our
society might have in future:
– Artifacts that are equal to animals
– Artifacts that are equal to humans (or at least appear like that)
– Artifacts that are superior to humans (post-singularity beings)
The first one can already be observed. Ranging from “tamagotchis” to Sony’s
Aibo robot-dog people bond to definitely inanimate tinboxes, more intelligent
artificial animals will cause even more passionate behavior. Nothing wrong with
that, if an artifact in deed has the intellect of a dog, why should it not receive
the same respect as a real dog? Machines equal to us is more of a challenge.
Phenomena like love, sympathy, racism or compassion might need a review if we
live door-to-door with these creatures. And they might discuss the very same
problems. The ultimate gain that we could get out of such a situation is to
overcome all antique concepts of difference and realize that conscious beings are
all equal. Anyway, we still did not encounter something entirely new. Throughout
the history of our planet we met equal cultures and had to cope with it.
Meeting an outclassing entity, however, is something that no-body is prepared for. Antique mythology tells of gods which are equipped with extraordinary powers but are still helplessly extradited to very human emotions like love
or anger. They can even be outsmarted sometimes, so they are by no means
almighty. Modern religions paint a picture of a supreme creator, immune to our
weaknesses, but captured in his own goodness and love for us. It is superior but
not free. A mixture – an autonomous, superior being – threatens us. Would they
exterminate us like we treat bugs and bacteria or would they explain us – finally
– the true meaning of life? At least the question of social implications would not
lay on our shoulders anymore, our superior creatures would ponder about that.
9 Conclusion and Outlook
Objects and devices get smarter every single day. It is a win-win situation for
the users – who can get more services by machines – and manufacturers who
always look for new products and markets. Their smartness today is shown in
terms of function, usability, design, energy efficiency, sustainable product life
cycle and the like. All of these are observable behavior, none of them can be
directly assessed in terms of directly influences. With conscious machines on
the other hand it would be possible to demand e.g. energy efficiency or security
for children. Machine consciousness when developed following the template of
human consciousness needs to show degrees of itself. It needs to be designed
in the way it could potentially work in a real, living body in its mediator role
20
Peter Palensky et al.
between endogenous demands and the actual environment. These functions are
elaborated by Freud and his successors who developed a functional model of the
human mind. Another attribute of this model lies in the concept of the primary
and secondary process and its implications for higher levels. Other humanities do
not possess functional models or provide only behavioral models. However, psychoanalysis turned out to be applicable to surprisingly large extent into technical
terms and in the future into systems.
The development of machine consciousness relys on many interdisciplinary
findings presented above, whereby computer engineering and psychoanalysis will
be the main contributors. Some of the requirements are already formulated in
concepts, some still lack any idea for implementation. However, we tried to give
the interested reader an impression of what has to be done to achieve machine
consciousness in our view. We also have stressed the borders: that even very
sophisticated solutions for human-like behaviour in terms of moving arms and
feet, producing facial expressions, following speech dialogs, etc. do not contribute
to making machines more conscious. The key for any stage of development of
higher order mental content lies in subjective evaluation of lower level mental
content. This is what children need to do from very soon after birth, and so
conscious machines will have to. Some remakable capabilities resulting from this
developmental process in children have been presented.
There is already a community, founded with the origination of the ENF –
the first international engineering and neuro-psychoanalysis forum [40]. Many
international researchers took notice of that event and even more than 100 came
to attend. Things start to come together. We see this as radically new approach
in AI. Many other approaches to reach human-like intelligence failed because of –
in retrospect – clearly visible lacks in theory or methodology. This one is unique
in terms of understanding and applying proven theories of human consciousness
into machines. You are welcome to join us and enhance our efforts with your
valuable ideas.
References
1. Baars, B. J.: Some essential differences between consciousness and attention, perception, and working memory. Consciousness and Cognition, vol. 6, 363-371 (1997)
2. Damasio, A. R.: The Feeling of What Happens: Body, Emotions and the Making of
Consciousness. Econ Ullstein List Verlag GmbH (1999)
3. Barnard, E., Palensky B., Palensky P.: Towards Learning 2.0. Proceedings of ICST
IT-Revolutions 2008, Venice (2008)
4. Palensky, P., Lorenz, B, Clarici, A.: Cognitive and Affective Automation: Machines
Using the Psychoanalytic Model of the Human Mind. Proceedings of First IEEE
Engineering and Neuro-Psychoanalysis Forum, Vienna (2007)
5. Arlow, J.A., Brenner, C.: Psychoanalytic Concepts and the Structural Theory. New
York: International Universities Press (1964)
6. Bateman, A., Holmes, J.: Introduction to Psychoanalysis – Contemporary Theory
and Practice. London and New York: Routledge. (1995)
AI 2.0
21
7. Chalmers, D.: The Puzzle of Conscious Experience. Scientific American, Dec. 1995,
62-68 (1995)
8. Chalmers, D.: Facing up to the problem of consciousness. Journal of Consciousness
Studies 2 (3), 200-219 (1995)
9. Dennett, D.: Who’s on First? Heterophenomenology Explained. Journal of Consciousness Studies, Special Issue: Trusting the Subject? (Part1), 10, No. 9-10, Oct.
2003, 19-30 (2003)
10. Sandler, J., Holder, A., Dare, C. & Dreher, A.U.: Freud’s Models of the Mind. An
Introduction. London: Karnac. (1997)
11. Freud, S.: Formulations on the Two Principles of Mental Functioning. In: J. Strachey (Ed. & Trans.) The Standard Edition of the Complete Psychological Works of
Sigmund Freud (Vol. 12, pp. 218-226). London: Hogarth Press (1911)
12. Freud, S.: The Ego and the Id. Standard Edition, Vol. XIX, 109-121 (1923)
13. Freud, S.: The Unconscious. Standard Edition, Vol. XIV, 166-204 (1915)
14. Freud, S.: The Interpretation of Dreams. Standard Edition, Vol. IV & V (1900)
15. Freud, S.: Three essays on the theory of sexuality. Standard Edition, Vol. VII,
135-243 (1905)
16. Freud, S.: Beyond the pleasure principle. Standard Edition, Vol. XVIII, 7-64 (1920)
17. Solms, M. & Turnbull, O.: The Brain and the Inner World. London: Karnac (2002)
18. Solms, M.: Freud returns. Scientific American, May, 56-62 (2004)
19. Solms, M.: What is the “mind”? A neuro-psychoanalytical approach. In: Dietrich, Zucker, Bruckner, Fodor (Editors): Simulating the mind, A technical, neuropsychoanalytical approach. SpringerWienNewYork. (2008)
20. Viola P., Jones, M.: Rapid Object Detection using a Boosted Cascade of Simple
Features, Conference on Computer Vision and Pattern Recognition (2001)
21. Mattern, F.: Ubiquitous Computing: Schlaue Alltagsgegenstände - Die Vision von
der Informatisierung des Alltags. In: Bulletin EV/VSE, Nr. 19, 9-13 (2004)
22. Hainich, R. R.: The End of Hardware, A Novel approach to Augmented Reality.
BookSurge Publishing (2006)
23. Lindwer, M., Marculescu, D., Basten, T., Zimmermann, R., Marculescu, R., Jung,
S. and Cantatore, E.: Ambient Intelligence Visions and Achievements; Linking abstract ideas to real-world concepts. Proceedings of the conference on Design, Automation and Test in Europe DATE03. (2003)
24. Endres, C., Butz, A. and MacWilliams, A.: A Survey of Software Infrastructures
and Frameworks for Ubiquitous Computing. In: Mobile Information Systems Journal
1, Nr. 1 (2005)
25. Lagendijk, R. L.: The TU-Delft Research Program “Ubiquitous Communications”
Proceedings of the Twenty-first Symposium on Information Theory in the Benelux,
33-44 (2000)
26. Roman, M., Hess, C. K., Cerqueira, R., Ranganathan, A., Campbell, R. H. and
Nahrstedt, K.: Gaia: A Middleware Infrastructure to Enable Active Spaces. IEEE
Pervasive Computing, 74-83 (2002)
27. Mavrommati, I. and Kameras, A.: The evolution of objects into Hyper-objects.
Personal and Ubiquitous Computing 7, Nr. 1, 176-181 (2003)
28. Gellersen, H-W., Schmidt, A. and Beigl, M.: Multi-Sensor Context-Awareness in
Mobile Devices and Smart Artefacts. Mobile Networks and Applications 7, 341-351
(2002)
29. Lohse, M. and Slusallek, P.: Middleware Support for Seamless Multimedia Home
Entertainment for Mobile Users and Heterogeneous Environments, 217-222 (2003)
22
Peter Palensky et al.
30. Want, R., Schilit, B., Adams, N., Gold, R., Petersen, K., Ellis, J., Goldberg, D.
and Weiser, M.: The PARCTAB ubiquitous computing experiment. Proceedings of
the Fourth Workshop on Workstation Operating Systems (1995)
31. Bion, W.: A theory of thinking. International Journal of Psycho-Analysis (43), 4-5
(1962)
32. Pfeifer, R., Scheier, C.: Understanding Intelligence, MIT Press (2001)
33. Brooks, R.A.: Intelligence without representation. iss. 47, 139-159 (1991)
34. Brooks, R.A.: I, Rodney Brooks, Am a Robot. IEEE Spectrum 06.08, (2008)
35. Alter, T.: Qualia. in Nadel, L. (Ed.): Encyclopedia of Cognitive Science. London:
Macmillan Publishers Ltd., 807-13 (2003)
36. Bringsjord, S.: The Zombie Attack on the Computational Conception of Mind.
Philosophy and Phenomenological Research 59.1 (1997)
37. Holland, O. (ed): Machine Consciousness. Imprint Academic (2003)
38. Penrose, R.: The emperor’s new mind. Oxford University Press (1989)
39. Fonagy, P., Target M., Gergely, G., Jurist, E.L.: Affect Regulation, Mentalization,
and the Development of Self. Other Press, 1 edition (2000)
40. Dietrich, D., Fodor, G, Zucker, G., Bruckner, D. (eds.): Simulating the Mind.
Springer (2008)
41. Förster, Heinz von: Wissen und Gewissen: Versuch einer Brücke, 7. ed. (2006)
42. Yowell, Yoram: Return of the zombie - Neuropsychoanalysis, consciousness, and
the engineering of psychic functions. In: Dietrich, Zucker, Bruckner, Fodor (Editors):
Simulating the mind, A technical, neuro-psychoanalytical approach. SpringerWienNewYork. (2008)
43. Beck, Friedrich and Eccles, John. C.: Quantum aspects of brain activity and the
role of consciousness. Proceedings of the National Academy of Sciences of the United
States of America, Vol. 89, p.11357-11361 (1992)
44. Roger Penrose and Stuart Hameroff: Orchestrated objective reduction of quantum
coherence in brain microtubules: The “orch OR” model for consciousness. Mathematics and Computers in Simulation 40:453-480 (1996)